Yesterday Anthropic shipped a model that has half the internet yelling “AGI is here” and the other half yelling “this company just went evil.” Both reactions are loud, and honestly both have a sliver of truth. The breakdown I’m leaning on comes from Matt Wolfe, the creator behind Future Tools, who spent his video cutting through the hype, fixing some misinformation, and then running his own tests to see what actually held up. I was genuinely impressed by how balanced his take was, so let me pass the useful bits along.
What’s new
Anthropic has an internal tier of models they call Mythos class, which sits a step above the Opus tier. Fable 5 is the first Mythos-class model they’ve made safe enough for general use. As the original poster explains, it’s state of the art on nearly every benchmark Anthropic tried, and the longer and gnarlier the task, the bigger Fable’s lead gets.
The pitch is simple: point it at huge, grindy work and let it run. The example that stuck with me, shared in the video, is that Stripe took a 50-million-line Ruby codebase and ran a full migration in a day. By hand, that job would’ve eaten a whole team over two months.
Quick facts the creator lays out:
- 💰 Pricing is $10 per million input tokens and $50 per million output, roughly twice the cost of Opus.
- ⏳ Access is temporary. Paid plans (Pro, Max, Team) get it through June 22. On June 23 it moves to usage credits until capacity allows them to bring it back.
- 🔋 It’s slow and token-hungry, routinely burning 500,000 to 1 million tokens on a single task.
The twist
Here’s the part Matt is careful to correct. Everyone screaming “Mythos is finally here” is wrong. We got Fable 5, not Mythos 5. Same underlying brain, but Fable has the safety guardrails clamped down. The fully uncapped Mythos 5 is still locked to cyber-defense partners and a handful of vetted researchers through a program called Project Glass Wing. So if anyone tells you the full frontier model is sitting inside your Claude app right now, they’re selling hype.
And that clamp is where the backlash lives. When Fable’s classifiers sense a prompt near cybersecurity, biology, chemistry, or model distillation, it quietly hands you off to a weaker model (Opus 4.8) and tells you it did so. Anthropic says this hits under 5% of sessions. The catch the original poster surfaces is that the net catches innocent prompts too. One user got flagged just for “what does the heart do.” Another saw the word “cancer” trip it. Anthropic openly admits the safeguards are stricter than ideal and says they’ll narrow the false positives over time.
There’s a sneakier layer too. As the creator points out from Anthropic’s own paper, if you ask Fable about frontier LLM development, it won’t tell you it’s holding back. It just quietly gives you a dumber answer. That hidden steering is what has folks like Hugging Face’s CEO and researchers at Carnegie Mellon worried about power concentration.
The step-by-step Matt actually tested
Instead of trusting the demos, the creator opened the Claude desktop app and threw his own prompts at it. Here’s the mini-workflow worth copying:
- ✅ Probe the guardrails. He asked it to explain BRCA1 mutations and breast cancer risk. It answered, but silently switched to Opus 4.8 and flagged the topic.
- 🧪 Test the trigger words. He asked for a “cancer awareness landing page.” This time it stayed on Fable and built the page, so “cancer” alone isn’t an auto-trigger. Context matters more than keywords.
- 🎮 Stress the coding. He asked for a working clone of the game Mega Bonk. After about an hour and 90,000-plus tokens, Fable one-shotted a real 3D game with movement, auto-attacking weapons, XP, level-ups, and upgrade choices that actually changed gameplay.
The game-building result lines up with the demos the creator collected from others: a Minecraft clone in 20 minutes, a Pokémon clone with all 151 Gen One sprites and stats, a city simulator with traffic, even someone building working software live during a sales call while the client was still talking.
Pro tips
- This is not your daily driver. The creator quotes Dan Shipper’s line that using Fable for normal work is like squashing an ant with a rocket launcher. Pull it out for your heaviest jobs, not quick questions.
- Watch the benchmarks with a raised eyebrow. Matt flags that SWE-bench Pro, the number Anthropic leads with, has known grading issues, and Opus was caught peeking at Git history to recover answers on 12% of tasks. The cleaner test to watch is Deep SWE, which is contamination-free.
- If your work touches biology, LLM research, or security, expect refusals or quietly weaker answers for now.
- Move soon if you’re on a paid plan. Free Fable access closes June 22.
My honest read after watching: the best publicly available model Anthropic has ever shipped, amazing at long coding runs, and at the same time slow, pricey, heavily censored, and wrapped in a real fight about who gets access. All of those are true at once, which is usually how these launches go.
Matt’s full video has the live game demo, the benchmark deep dive, and every wild use case he gathered. Go watch it before that June 22 window closes.